Views
No views yet
.ply/.splat), MIT. The zoo's first 3D model: outputs drop straight into a Gaussian-splat
viewer (e.g. Apple RealityKit on visionOS, or MetalSplatter on iOS/macOS)..aimodel bundles (each is a directory). Conversion + runner scripts
live in the coreai-models-community zoo (conversion/triposplat/).apps/TripoSplatMac,
the zoo app that ships this model (single image → 3D Gaussian splats on Mac: 5 converted nets + host flow sampler and octree resampling; build & run steps in its README).| net | shape | bundle | dtype |
|---|---|---|---|
| DINOv3 ViT-H encoder | (1,3,1024,1024)→(1,4101,1280) | dinov3_fp16.aimodel | fp16 |
| Flux2-VAE encoder | (1,3,1024,1024)→(1,4096,128) | vae_fp16.aimodel | fp16 |
| DiT denoiser (one step) | latent(1,8192,16)+cam(1,1,5)+t+feat1(1,4101,1280)+feat2(1,4101,128)→latent,cam | dit_fp16.aimodel | fp16 |
| Octree probability decoder | x(1,8192,3)+l(1,)+cond(1,8192,16)→logits(1,8192,8) | octree_fp32.aimodel | fp32 |
| Decode (gs + build_gaussians + .ply activations, baked) | points(1,8192,3)+cond(1,8192,16)→(262144,14) | decode_fp32.aimodel | fp32 |
FlowEulerCfgSampler) and the octree sample_probs systematic resampling
stay host-side (data-dependent control flow). Scripts: _conv_*.py convert+gate each net;
_conv_fp16.py makes the half-size fp16 bundles; _conv_decode.py bakes build_gaussians + the
Gaussian .ply-activation math into one net so the runner just writes raw floats.model.py; all are general gotchas:aten.arange → bad_optional_access C++ abort. Use int-arg arange (DINOv3 RoPE).got multiple values for 'mod' — submodule called with mod= kwarg. Pass positionally.torch.polar/view_as_complex) as real
cos/sin math (apply_rotary_emb, RePo3DRotaryEmbedding.forward).sin/cos of huge args is low-precision (cos→0.5) — the DiT positional embed
computed from the fixed Sobol constant was folded wrong; precompute it into a register_buffer.F.normalize drops the eps clamp → near-zero vectors blow up ~1e13; rewrote MultiHeadRMSNorm
as explicit x*rsqrt(mean(x²)+eps). (Emergent only at large seq len — gate by VISUAL/true-scale.)prog.optimize() hangs on the 24-block/12k-token DiT graph (>90 min) — skip it
(convert(optimize=False)), AOT coreai-build optimizes for the device anyway.l (resolution) input → CoreAIError 3 at runtime, pass it as float32._run_coreai.py (or app_backend.py --input <img>) loads the bundles via coreai.runtime
(SpecializationOptions.default() = GPU; ~2 min/gen at 20 steps on Apple silicon, full quality).
End-to-end latent gate vs torch-DiT: cos 0.999999.TripoSplatMac (standalone) and TripoSplatPhone
(capture on iPhone → Mac server server.py → view splats in MetalSplatter / RealityKit)..aimodelc
is ~3.1 GB and the DiT's 12294-token full-attention score matrix alone is ~4.8 GB, both over the
~3.3 GB iOS app memory budget (weight precision doesn't fix the attention working set). Needs
flash-attention conversion / weight streaming. The Mac-link client is the shipped path.